Symposium on Del Pinal and Spaulding, “Conceptual Centrality and Implicit Bias”

I’m very glad to announce our latest Mind & Language symposium on Guillermo Del Pinal and Shannon Spaulding’s Conceptual Centrality and Implicit Bias” from the journal’s February 2018 issue. Commentators on the target article include Bryce Huebner (Georgetown), Edouard Machery (Pittsburgh), Eric Mandelbaum (CUNY), Steven Sloman (Brown), and Ema Sullivan-Bissett (Birmingham).


The term “implicit bias” is typically used refer to unconscious or introspectively recessive attitudes, associations, or stereotypes that can prejudicially affect our judgments, decisions, and actions with respect to members of a social category — for example, Asian Americans, gay men, or Jews — and, moreover, do so in way that is automatic and difficult intentionally to control. This definition is representative, but not uncontentious. Among other things, there is debate concerning the extent to which implicit biases are genuinely inaccessible to consciousness and whether psychological measures of implicit bias are predictive of discriminatory or otherwise morally undesirable behavior in real-world contexts. (Check out the Brains Blog Roundtable: What can we learn from the Implicit Association Test?) The definition also leaves open important questions concerning the format(s) of the mental representations that encode implicit biases, the underlying unity or “psychological kindhood” of implicit bias phenomena, and the proper way to draw the explicit/implicit distinction (for a diverse range of philosophical perspectives, see the essays collected in Brownstein & Saul 2016 a,b.) An important merit of both the target article and the commentaries in the current symposium is that they help elucidate the nature of implicit bias and collectively take steps toward answering all of these questions.

In the target article, Del Pinal and Spaulding suggest that philosophical and empirical discussions of implicit bias have focused on salient or statistical associations between target features and representations of social categories at the expense of examining other ways in which implicit biases might be encoded. In particular, they argue that recent discussions of implicit bias have overlooked the dependency networks that are part of our representations of social categories. Dependency networks are structures that capture information about how features in a conceptual representation depend on each other, which in turn determines their degree of centralityTo estimate the degree to which a feature f is central to a concept C is to estimate the extent to which others features encoded in C depend on f (Sloman et al. 1998). One especially relevant type of dependence is causal. Other things beings equal, a feature f will be more central to C the more it is a cause rather than an effect of other features (Ahn et al. 2000; Rehder & Kim 2006, 2010; Sloman 2005; Sloman & Lagnado 2015). A concept C, then, can encode implicit bias in either of two ways: the bias can result from the salient-statistical properties of the C’s constituent features or it can result from the dependency networks that link those features.

Del Pinal and Spaulding argue that centrally encoded biases systematically disassociate from those encoded in salient‐statistical associations: a feature f could be relatively central in C while not having either high cue validity/saliency and, going in the other direction, f could have high cue validity/saliency for C while not being relatively central (Section 3). A feature’s degree of centrality, they further maintain, determines its cross‐contextual stability (Section 5). In other words, the more central a feature is for a concept, the more likely it is to survive into a wide array of social categorization and induction tasks involving that concept. 

The distinction between implicit biases supported by salient or statistical associations and those supported by centrally encoded biases, Del Pinal and Spaulding conclude, has significant implications for debates concerning the underlying unity and representational format of implicit biases (Section 6):

…even if we focus just on the class of biases that involve relations between concepts and cognitive features, implicit biases can be encoded in different ways…. It follows that we should not assume that this broad class of implicit biases has a uniform underlying nature. Thus, when investigating whether implicit biases are beliefs, we should examine each case independently. For, not only may the answer be different for different kinds of implicit bias, each bias may correspond to a different form of belief (108).

…theorists concerned with the metaphysical status of biases should also investigate questions about their degree of centrality. Particular kinds of sensitivity to evidence are important factors in deciding whether a mental state counts as a belief of a certain form. To carry out the relevant investigations, we need different kinds of experiments than those that have dominated the implicit bias literature (109).



Ahn, Woo-Kyoung, & Nancy S. Kim (2000). “The causal status effect in categorization: An overview.” Psychology of learning and motivation. Vol. 40. Academic Press, pp. 23-65.

Brownstein, M. & Saul, J. (2016a). Implicit bias and philosophy volume 1: Metaphysics and epistemology. New York, NY: Oxford University Press.

Brownstein, M. & Saul, J. (2016b). Implicit bias and philosophy volume 2: Moral responsibility, structural injustice, and ethics. New York, NY: Oxford University Press.

Rehder, B., & Kim, S. (2010). “Causal status and coherence in causal-based categorization.” Journal of Experimental Psychology: Learning, Memory, and Cognition, 36(5), 1171-1206.

Sloman, S. A., Love, B. C., & Ahn, W. K. (1998). “Feature centrality and conceptual coherence.” Cognitive Science22(2), 189-228.

Sloman, S. (2005). Causal models: How people think about the world and its alternatives. NY: Oxford University Press.

Sloman, S. A. & Lagnado, D. (2015). “Causality in thought.” Annual Review of Psychology, 66, 223–247.


Comments on this post will be open for at least a couple of weeks. Many thanks to Guillermo, Shannon, and our five superb commentators. All of us here at the Brains Blog are also grateful to Gregory Currie, the other Mind & Language editors, and the staff at Wiley-Blackwell for their continued support of these symposia.

  • You can learn more about Guillermo Del Pinal and his research here.
  • You can learn more about Shannon Spaulding and her research here.

Below there are links to the authors’ introduction, the target article, commentaries, and the authors’ replies.

Guillermo Del Pinal and Shannon Spaulding: Introduction

Target Article: “Conceptual Centrality and Implicit Bias”

Commentaries and Replies:







  1. Eric

    In response to Sullivan-Bissett DPS writes “The objection, due to Madva (2016), is that if implicit biases were beliefs, they should be sensitive to logical form; but there is evidence that at least some aren’t.” DPS like this argument, and think it’s the strongest against ESB. But as a doxasticist I wouldn’t worry much about this argument. If it worked it would show that we have no beliefs whatsoever. Every belief (or state for that matter) sometimes fails to function the way it’s supposed to. In fact, the more strongly held the belief is, the less frequently it will function the way it “should.” I use scare quotes because there are multiple laws that hold over belief states. Some of these predict that strongly held beliefs will increase in strength when one receives disconfirming evidence (consider a case where I give you some evidence that you are a morally bad person—it’s highly unlikely the logical form of the counterevidence will just cause you to change your belief). I think belief perseverance and belief polarization aren’t mere performance constraints, but inform the underlying laws of belief (see my “Troubles with Bayesianism: An Introduction to the Psychological Immune System” for extended arguments). But you don’t have to believe in the nomological part to believe in the existence of belief perseverance–you can reasonably believe instead it is just a performance/competence issue. And belief perseverance is about *belief.* Of course, there’s nothing special about belief here–any state may fail to update because of (e.g.,) inattention, or competing demands. The interesting data (at least here) is about the competence of the state–it’s that it can update merely based on logical form (this was one of the main points I tried to get across in “Attitude, Inference, Association”). The fact that it sometimes fails to isn’t itself evidence against its being a belief.

    While I’m here I might as well reiterate an earlier question I mentioned that seems like it went unanswered (though I may have missed it): what do DPS take the difference between implicit attitudes and implicit biases to be? Honestly curious (and have no particular answer myself).

  2. Hi Eric!

    Thanks for your comments. You’re right that some beliefs are not sensitive to putative countervailing evidence, and this does not impugn their belief status. (We’ll definitely check out your “Troubles…” paper, as we are currently working on a paper on implicit bias and beliefs.) In our comments, we addressed a different problem with this particular argument against doxasticism. We argued that the experimental setup (saying “no” to a stereotype vs. “yes” to a counterstereotype) is not helpful in investigating whether implicit biases are beliefs because it doesn’t tell us much about the logical form of the representation. One might object to the argument for other reasons, too. (Indeed, in our conversations about the commentaries, we talked about the issue that you raise above. In the end we didn’t include this in our replies for the sake of space, so we’re glad you brought it up here.) So, of the arguments ESB presents against doxasticism, this does seem to be one of the stronger ones, but we agree that it is far from decisive.

    On the last question regarding the difference between implicit attitudes and implicit biases, strictly speaking we don’t have a view on whether all/some/no implicit biases are implicit attitudes. Some, like Machery (2016) and Huebner (2016), argue that we should not conceive of implicit biases as attitudes but as dispositional traits. That’s fine. For us, talking about implicit biases as attitudes (or beliefs) is just a convenient way to talk about implicit biases. One can have a debate about the nature of attitudes (or beliefs) and whether all/some/no implicit biases behave like attitudes (or beliefs). However, we think it is more fruitful to investigate the cognitive structures of the underlying representations. That will tell us more about their stability, sensitivity to evidence, inductive potential, etc.

    Huebner, Bryce. 2016. “Implicit Bias, Reinforcement Learning, and Scaffolded Moral Cognition.” In Implicit Bias and Philosophy, edited by Jennifer Saul Michael Brownstein. New York: Oxford University Press.

    Machery, E. 2016. “De-Freuding implicit attitudes.” In Implicit Bias & Philosophy, edited by M. Brownstein and J. Saul. Oxford: Oxford University Press.

  3. Eric

    Hey Shannon,

    Cool. Thanks for clearing that up for me. I very much agree with this: “We argued that the experimental setup (saying “no” to a stereotype vs. “yes” to a counterstereotype) is not helpful in investigating whether implicit biases are beliefs because it doesn’t tell us much about the logical form of the representation.” (Ditto for this: “However, we think it is more fruitful to investigate the cognitive structures of the underlying representations. That will tell us more about their stability, sensitivity to evidence, inductive potential, etc.”)

  4. The comments and replies are super interesting and informative! A ton to digest here so I’m glad comments will be open for a few weeks!

    Just for the record — I think there’s a ton of important stuff to talk about here beyond the “traditional” associative vs. propositional debate, and I plan to say more in future posts — the way I tried to articulate responsiveness to logical form in that 2016 paper was meant to allow for all sorts of deviance, cognitive-dissonance shenanigans, the performance/competence distinction, etc. Whether I succeeded in doing so is another matter. Also just for the record, one of the central arguments of that paper was that the existing evidence is extremely inconclusive in lots of ways (including the “no” to stereotypes stuff, but also including the evidence assembled by Eric, Neil Levy, and Jan de Houwer and colleagues), and to point toward all sorts of other kinds of empirical studies that could be informative. (And the relevant research that’s come out since that paper has, I think, made things look even more complex, messy, and weird than they did before, but that’s a topic for another time…)

    However, I’d like to re-emphasize some important contributions from Guie and Shannon’s paper that could easily be overlooked if we try to “force” their paper into one side of the existing debates about the cognitive structure of implicit bias. That debate has largely played out in terms of whether “associative” or “propositional” structures (or processes, etc.) explain performance on timed association tasks like the IAT or priming measures like EPT and AMP. However, a bias need not manifest on any of the leading implicit measures (or explicit measures, such as feeling thermometers or Likert-scale agreements with statements of various kinds) in order to be real and causally significant. Instead, Guie and Shannon propose that there may be non-associative biases best targeted by *non-associative* measures. Someone could easily maintain that the biases measured by the IAT and EPT are encoded as salient-statistical associations but that there are all sorts of other more robustly structured phenomena (such as dependency networks) that deserve to be called, or connected to, “implicit biases,” and which are *not* best measured in those ways. Of course, even this would be somewhat of an oversimplification, because it is widely acknowledged that the IAT is not “process-pure;” performance reflects a bunch of difference processes. So one could also take the view that both salient-statistical associations and dependency networks are involved in IAT performance. There are lots of possibilities to explore here… Perhaps the dependency networks should go in the bucket with “explicit propositional processes” alongside cognitive consistency, such that both play a role in defining the structure of reasoning. (There’s a whole other can of worms here about why to think of centrality as implicit which I hope to address in a future post.) Anyway, Sloman’s comments raised some really important and difficult questions about this set of issues, and I wouldn’t want this point to get overlooked in the ensuing discussion.

    The second important set of contributions has to do with moving beyond the debate of “do these biases have some structure or virtually no structure at all?” to *which specific structure(s) do they have*. Different people could agree with the statement “Muslims are terrorists” or associate “Muslim” and “terror” on an IAT but have very different cognitive structures implicated. It could be: given that so-and-so is a Muslim, there is a higher likelihood that he is a terrorist; or given that so-and-so is a terrorist, there is a higher likelihood that he is a Muslim; or it could be something else entirely. For example, I suspect centrality, rather than sheer probabilities, salience, or associations, is involved when anti-Muslim rhetoric alludes to the calls for jihad in the Qur’an. The idea is that there is some “deep,” internal feature of the religion which explains the violence. The alleged central role of violent jihad *can* explain the alleged prevalence of violence or terrorism or aggression or whatever in the Muslim community (and so we *should* expect that it relates to various other sorts of implicit and explicit measures in complex ways), but it’s *also* a cognitive structure tailor-made to be resistant to apparent counter-evidence (pointing to all the non-violent Muslims or violent non-Muslims, etc.).

  5. Hi Alex,

    Thanks for your comments! You’re right that your 2016 paper (cited by Sullivan-Bissett) is cautious with respect to the relevant empirical evidence. We encourage folks interested in the topic to read that paper in full. And for those interested in the weird, complicating research that has come out since that paper was published, see Johnson et a. (2018). (Citations below.)

    As you note, our framework is neutral with respect to whether implicit biases are beliefs. For us, the most important issue is figuring out the underlying cognitive structures of implicit biases. And, as you also note, this is quite complicated! Sloman and Machery’s comments on the relation between centrally encoded biases and salient-statistical associations highlight just how tricky this issue is. One thing we are especially keen to work on is how to distinguish the context dependence of centrally encoded biases and the context dependence of salient-statistical associations. Though we do not endorse psychological essentialism in the paper, we maintain that centrally encoded biases are much more resilient across contexts than salient-statistical associations. If one endorses psychological essentialism, it is quite easy to see how centrally encoded biases could be more resilient, and we think there’s a plausible story to tell about how they can also be (somewhat) sensitive to context. If one rejects psychological essentialism and indeed rejects the idea that there are significant “centers of gravity” in feature dependency networks, explaining the context sensitivity of centrally encoded biases is simple. But then the task is to explain the distinctive resilience across contexts (e.g., subcategorization and inductive inferences) that we describe in the paper. We also think there’s a story to tell here. So, one aspect of Sloman and Machery’s comments we’re working on is how neutral the framework can be with respect to these background theoretical commitments. We are happy to take a stand on some of these issues if we need to, though we want our framework to be as general as it can be while also being useful for philosophical and empirical investigation of implicit biases. So, I suppose these remarks serve to echo what you say above, namely, that the replies raise new, interesting and difficult issues with respect to implicit bias research. We take this to be a very good sign of the fruitfulness of this approach to implicit bias and are incredibly grateful to our commentators for their thoughtful replies!

    Johnson, I. R., Kopp, B. M., & Petty, R. E. (2018). Just say no!(and mean it): Meaningful negation as a tool to modify automatic racial attitudes. Group Processes & Intergroup Relations, 21(1), 88-110.

    Madva, Alex. “Why implicit attitudes are (probably) not beliefs.” Synthese 193.8 (2016): 2659-2684.

    • Hi Shannon,

      I take the point that major “centers of gravity” in the network will be generally more context-independent. But I’m also pretty moved by Sloman and Machery’s suggestions that “more vs. less context-sensitive” is not necessarily the most perspicuous way to slice up the options, as opposed to something more fine-grained like “these sorts of typical features will be sensitive to these contextual factors but not those” and “these sorts of dependency relations will be sensitive to those contextual factors but not these.” It seems like your response to Machery on p.3 of your replies grants something like this point.

      Even if we acknowledge that there are “centers of gravity” (as I agree we should) there’s a further question about whether some of the most prevalent, pernicious, and causally significant social biases are actually located closer to these centers or not. Going with the men’s brilliance vs. women’s brilliance gender bias, it seems to me that we can make important contrastive claims like “people take women’s brilliance to rely more on hard work than innate talent” and “people take men’s brilliance to depend more than women’s brilliance on innate talent,” without making any commitments to just how “central” vs. “peripheral” these dependency relations are. There can be all sorts of important, causally significant asymmetric dependency biases located closer to the periphery than to the center. Their explanatory and social importance doesn’t stand or fall with their degree of centrality.

      I also want to think more about the extent to which these biases are implicit. I gather from your response to Sloman (pp.3-4) on this point that you’re thinking of these as “implicit” along the lines of Chomskyian tacit knowledge of grammar. We can ask people if they think such-and-such sentence is grammatical, which gives us information about their tacit knowledge, even though they perhaps cannot articulate the rules they follow, or even articulate rules that they don’t actually follow. So putative “opacity to introspection” would be what unites the structures implicated in IATs with the structures implicated in the centrality measures (maybe we should call centrally encoded biases “tacit” rather than “implicit”?). Still, I think it’s not clear yet, based on the data we have, whether (or perhaps how many of) the dependency biases we’re interested in meet these criteria. We haven’t done the tests yet, although the experiments would seem relatively straightforward. I think the strongest case for their being implicit (or explicit, depending where the chips fall) would be if we could throw together some empirical evidence demonstrating 1) that many people do not explicitly agree with statements like “women are just as likely to be smart as men, but men’s smartness depends on innate intellect and women’s smartness depends on hard work,” paired with evidence demonstrating 2) that many of these same people nevertheless also reveal centrality biases on other measures.

      Part of what I’m circling around here, however, is that I don’t think the implicit biases measured by the IAT are unconscious either. Here I’m following work that’s come out of several independent labs (Payne’s, Gawronski’s, Nosek’s, and others), suggesting that whatever these tasks are measuring is no *less* conscious than explicit attitudes. (Maybe all of our attitudes are unconscious, all self-knowledge is inferential, etc., but so-called implicit biases are not more unconscious than so-called explicit.) So one of the most promising ways of drawing the implicit/explicit distinction is just at the level of *measures*, which is that the implicit measures don’t depend on self-report and the explicit measures do. And it seems like the various promising centrality measures will differ with respect to this issue, some being more self-report-based and some being less so, or much more indirectly so, and some of them will be hard to place along this dimension altogether. Which brings me back to thinking that the strongest case for some kind of implicit-ness would be dissociations between more and less direct ways of getting at it. And it could be that self-presentation concerns or moral values prevent people from explicitly agreeing with statements like (1) above without their being unconscious. One of the central aspects of the aversive racism approach has been that people will be more likely to act on their biases when they don’t realize their biases are implicated in a given task, and I think that point’s at play in a lot of the centrality measures (unlike studies where you ask people if such-and-such sentence is grammatical, by the way, where perceptions of grammaticality are transparently being measured), and that issue is also independent of the opacity to introspection issue. (There can be attitudes completely transparent to introspection but that influence me in ways I don’t realize during a given task.)

      • Guillermo Del Pinal

        Hi Alex,

        Thanks for the very interesting comments! I want to add a few observations on the issues you raise concerning the relation between context-sensitivity, biases, dependency structures, and central vs peripheral encodings.

        It seems clear that, at this point, most of us agree that thinking in terms of dependency structures allows us to identify a class of biases that have to do with local dependencies, although they need not be central. Suppose that F1 uniquely depends on F2 in conception C1, whereas F1 uniquely depends on F3 in conception C2. We can predict that in contexts that in/directly drop/affect just F2, F1 will be more likely dropped/affected in C1 relative to C2, and vice-versa for contexts that in/directly affect just F3. These kinds of generalizations will be very useful in predicting the role of biases, once we uncover their local dependencies.

        If we generalize this kind of observation, it is tempting to think that, ultimately, all we really need to make predictions about the behavior of biases across contexts is to uncover the details of local dependencies. From this perspective, we might conclude that whatever rough heuristic value there was in saying something like “the more central the feature/bias, the more stable/resilient across contexts” is strictly speaking a kind of dispensable motto.

        However, it is worth exploring, briefly, precisely why we might not want to go so far…

        First, in some cases, it will be hard to uncover the details of local dependencies. When we studied the brilliance gender bias (in “Stereotypes, conceptual centrality and gender bias”), it was relatively easy to do this because our pre-studies and studies by other researchers gave us a good guess about what the key local dependencies could be (hardworking, disciplined, etc., in relation to smart, brilliant, etc.). In other cases, this might not be so easy to do, and so we might have to use the tasks which just trace the general mutability of a feature for a conception.

        Second, many of the ad hoc contexts we encounter in everyday life do not cleanly manipulate the variables uncovered by local dependencies. In these cases, having a measure of the overall centrality of a feature in a conception will continue to play an important role in predicting its behavior. This is mainly because, in these kinds of cases, we won’t be able to tell precisely what is happening to the key variables in the local dependencies.

        To illustrate, although given a context which adds information like `lazy Princeton fe/male professor’ we would be able to use (to predict the default behavior of the brilliance-gender bias) the local dependency of in each conception (`female’ vs `male Prof’), this will not be possible to do given contexts which add information like `coffee addict Princeton professor’, `Mexican raised Princeton professor’, `stand up comedian Princeton professor’, and so on. In general, we want to also predict how biases will behave in such superficially innocent ad hoc contexts, and in these cases having an overall measure of centrality is crucial.

        Third, we eventually want to determine whether there is an interesting generalization concerning the interaction between biases and types of social concepts along the following lines. Social concepts which approximate the structure of natural kinds (roughly, which have strong gravitational centers) will tend to have more biases dependent on those centers compared to social concepts which approximate the structure of artifactual kinds. The relevant contrast here, to a first approximation, might be found when we compare between our conceptions of ethnic and gender social concepts with those of, say, social class or political affiliation.

  6. Cool beans. Just to clarify, I didn’t mean to suggest that we should only be interested in local dependencies, or that more centrally located features will be unimportant, etc. I’m totally on board for mapping out the full conceptions as much as possible, both for its own sake and, e.g., for debiasing purposes, where it might be the case that changing some of the more centrally located features will have a broader range of valuable downstream effects.

    But I did mean to suggest that *some* of the features we’re most interested in (for social and cognitive explanation, related to social justice and more broadly) might not be located at the center. It’s an open question.

    Three more points related to this stuff…

    1. I take it that one of the upshots of research on intersectionality is that many of the traits we assume to be relatively central actually end up being more context-sensitive when you look at interacting social categories. So some of the apparently superficial transformations you have in mind might not be so superficial, at the level of conceptions. It’s an empirical question which are which. But for example, Pedulla (2014) found that participants offered a higher starting salary for a straight white male candidate over a straight black male candidate and over a gay white male candidate– but there was no salary penalty for a black gay man relative to a white straight man. Why might this be? Pedulla assembles some evidence to suggest that the stereotypes associated with “black male” and “gay male” somehow “cancel out” (aggressiveness and femininity, respectively). Other studies have found similar patterns with other intersections of kinds. I take it that, on a picture where there are certain central features that survive these transformations, the cancel-out story is hard to make sense of. I might be mistaken about that, but in any case, I wouldn’t want to give up on the idea that aggressiveness and femininity are perceived by participants not just to be “typical” features of these categories but to be categories that are tied to stereotypical “black essences” and “gay essences”, etc. I think dependency relations should be part of the story regardless of how centrally located they are.

    2. Related to #1, one mistake we want to avoid is overpredicting bias and discrimination, or supposing these phenomena to be, in the complex real world, more stable and predictable than they are. It might be that the real-world reasoning patterns and discriminatory behaviors that we want to explain are themselves noisy and mercurial and contextually variable, and our best explanations should reflect as much. (I think this point is important for thinking about the putatively profound inadequacies of the IAT as well.)

    3. One way that social psychologists and some philosophers (like Elizabeth Anderson) have talked about the transition from old-fashioned racism to more modern forms of racism is to suggest (to put it in the parlance here) that the typical features in people’s conceptions have remained the same while the dependency relations have changed. Specifically, old-fashioned racism is characterized by things like a belief that black people’s disadvantages, stereotypical behavioral traits, etc., are explained by their having an inferior (or violent or hypersexual, etc.) biological essence. The transition to more “modern” racism is to shift the explanation away from biological essence and toward things like “culture” and “value” and “irresponsible choices.” So this might be one sort of example where the typical features are resilient even while the features they depend on are thought to change. (And someone might even explain some of the putatively typical traits by appeal to structural oppression. Now is as good a time as any to advertise the work of Nadya Vasilyeva who has some work in prep finding that under some conditions children incline toward structural explanations of generics like “boys play with green balls”.) Of course, there will be much more to say about these sorts of case. I don’t mean to introduce them as decisive in anyway, just more to chew on.

    Pedulla, D. S. (2014). The Positive Consequences of Negative Stereotypes: Race, Sexual Orientation, and the Job Application Process. Social Psychology Quarterly, 77(1), 75–94.

Comments are closed.

Back to Top